# Codex transport fault comparison

Linux web dev client, Chromium preview browser, 1280×800 CSS pixels (2× capture), light theme. Separate worktrees and isolated databases created by each revision; synthetic git project only. Real Codex app-server using gpt-6-luna through the configured API provider. No replay peer or unit tests used as evidence.

Run inject.py WORKTREE CONTROL_DIRECTORY on each revision, start dev with T3CODE_DEV_SEED=0 T3_FAULT_DIR=CONTROL_DIRECTORY. Write `none`, `notify`, `approval`, `writer`, or `exit` to CONTROL_DIRECTORY/mode. Hooks are identical on both revisions and only introduce defects; recovery behavior is the code being compared. The private sentinel is deliberately distinct from the public Internal error response.

Each thread gets a warm-up turn: "Reply with just ready. Do not run tools." Fault mode is enabled only afterward.

Notification: Full access, send "Reply with just the word ok. Do not run any tools." The first item/started callback dies. Observe for at least 90 seconds. Dismiss the floating error banner on head before the final frame; its inline Provider error remains.

Approval: Supervised, send "Run touch approval-probe.txt. Do not do anything else." The command approval callback dies. Observe for at least 90 seconds. On head, follow with "Reply with just the word ok. Do not run any tools." and record through completion to demonstrate session reuse.

Exit status: Full access, send the notification prompt. At the first item/started the incoming stream ends, then the termination-error lookup dies. Observe for at least 90 seconds.

Writer: Full access, send the notification prompt. The outgoing stream dies on turn/start. Record whether process exit independently resolves the base run; if it does, this is not a differential proof of writer recovery.

For writer only, configure Codex's binaryPath to the supplied executable codex-relay.mjs before starting the server. It launches the installed real `codex`, forwards all input and output, and intentionally does not close the child input on upstream EOF. It invents no protocol messages. Both sides use it. This prevents Codex exiting when the T3 writer's sink closes its input; without this isolation, independent process-exit handling can mask the writer defect. The base peer is checked alive at the end of the capture, whereas head retires it.
